Quantification of Textile-Based Stretch Sensors Using Machine Learning: An Exploratory Study

摘要

Goal: Textile-based stretch sensors are a novel and innovative alternative to traditional wearable sensors with applications in many different fields including robotics, virtual reality and healthcare. However, due to their non-linear properties it can be challenging to obtain accurate information. The goal of this study was to investigate if machine learning can be applied to obtain more accurate measurements. Methods: In a tensile test using a linear stage setup, data were collected from two commercial available stretch sensors (Adafruit and Image SI) and one self-fabricated sensor (Menrva research group at Simon Fraser University, Canada). For each sensor, one hour of consecutive stretches in both a trapezoidal and sinusoidal input pattern were collected. We identified a set of features, trained three commonly used machine learning algorithms, and compared their performance in estimating the amount of stretch. To demonstrate the feasibility of our approach in real life, we tested our setup in two human applications. First, we attached a stretch sensor to the human chest to estimate the expansion of the rib cage during breathing. Second, we evaluated the performance in estimating the ankle position with a sensor attached to the foot. Results: In the tensile test, Support Vector Regression performed best with an average accuracy (R